How to use DDP + FP16 without fit?

[turian]

We have a model that doesn't require fit(). We just want to run inference with it using DDP + FP16.

Shoe-horning it into lightning is a little strange, but it gives use DDP + FP16 very easily. However, it appears that we need a dummy fit() (with dummy optimizer and train_step).

Without fit(), we get strange errors just invoking test. How can we use DDP + FP16 without fit?

[turian]

Seems to work?

[carmocca]

I'm not sure I get what you want

If what you want is to run test, you do not need to run fit before

model  = ...
trainer = Trainer(accelerator="ddp", precision=16)
trainer.test(model)

Should work with no issues